AI search is two games with different rules.

What are the two AI-search games?

Game one is earning visibility inside Google's search system. Game two is earning mentions and citations inside engines that assemble answers using their own retrieval. They overlap in what makes a page good. They differ in almost everything you would act on.

Why is Google AI still an SEO game?

Google states that a page must be indexed and eligible to appear in Search before it can be shown in AI Overviews or AI Mode. That makes crawl access, indexation, internal linking and ordinary usefulness foundational rather than optional legacy work.

This is the single most cost-saving fact on this page. The market sells files and markup as AI visibility deliverables. The platform whose features they claim to influence says in writing that none of it is required.

Why measure ChatGPT, Claude and Perplexity separately?

Because they read different things. ChatGPT runs a proprietary index alongside external search and vendor feeds. Claude searches through Brave, closely enough that Brave position is a usable diagnostic for whether Claude can find a site at all. Perplexity maintains its own index. The same prompt can therefore return three different shortlists, and none of them has to match Google.

Answers are also probabilistic and personalised. The same prompt run twice in fresh sessions can name different companies, which is why any honest measurement is a rate across repeated runs rather than a screenshot.

How little the two games overlap

The overlap is measurable and it is small.

• Ahrefs analysed 15,000 queries and found roughly 12% of URLs cited by AI tools also appeared in Google's top ten organic results.
• In a separate March 2026 study of 4 million AI Overview URLs across 863,000 result pages, 37.9% of cited URLs also appeared in the first ten blocks of the ordinary page.

Ahrefs sells SEO tooling, and a finding that rank and citation are diverging is not obviously in its commercial interest, which is worth noting in its favour rather than against it.

The practical consequence is the useful part. A page sitting at position 80 can still be the source an engine quotes, because the retrieval is not the same retrieval. Ranking first is neither necessary nor sufficient.

Google is two surfaces, not one

We have been treating "Google AI" as a single thing on this page. It is two, and a business can be in one and absent from the other.

AI Overviews fire automatically on the ordinary results page, cite roughly three domains, and barely appear on local intent, where the map pack still answers. AI Mode is a separate tab the user chooses, keeps conversation state, cites around seven domains, and leans on LinkedIn, YouTube and Reddit. Reported overlap between the two, for the same question, is well short of complete.

Since August 2026 an AI Overview can also auto-expand into a fuller, AI-Mode-style response, which is what we saw on our own query in September. So the surfaces are converging in appearance while still retrieving differently underneath. Audit them as two: same prompt set, logged separately, because the fix differs.

The crawler map

Search access and training consent are separate permissions, and most robots.txt files conflate them. Every row below was read from the platform's own documentation on 23 September 2026.

The mistake almost everyone makes

Blocking Google-Extended does not remove you from AI Overviews. It governs the Gemini apps and Vertex AI grounding. Google's AI features in Search are controlled by ordinary Googlebot rules and snippet directives such as nosnippet, data-nosnippet, max-snippet and noindex. Since 31 August 2026 there is also a Search generative AI control in Search Console, which takes a whole site out of AI Overviews, AI Mode and Discover's AI features while leaving it in ordinary Search.

Both errors are common and both are expensive. A publisher who blocks Google-Extended believing it opts them out of AI Overviews has given up Gemini grounding and kept the thing they objected to. A business advised that allowing it will help AI Overviews has been sold a change with no effect on that surface.

What do both games share?

Both reward a business that is easy to fetch, easy to understand, and independently corroborated. Technical health, focused pages, an answer near the top rather than buried, facts that agree across every profile, and evidence somebody else published. The mechanics of how an engine selects and cites a source are worth understanding before spending against either game, and markup is not the lever it is sold as.

One strong website can support every channel. One measurement model cannot describe every channel.

How should a business measure the two games?

1. Track impressions, positions, clicks and conversions for target pages in ordinary search, plus AI impressions in Search Console's generative AI report, which gives impressions only and no clicks or queries.
2. Build a fixed prompt set from real discovery and comparison questions, not from keywords.
3. Per engine, record four separate states: retrievable, retrieved, cited, and named in the text. Where we have run this, the gap has most often been between retrieved and cited. That is an observation from a small number of checks, not a measured rate, and it usually points to a passage problem rather than an authority problem.
4. Record what was cited instead of you. That list is the plan.
5. Re-run the same set monthly in fresh sessions and report rates, annotating the chart with update dates.

Limits

• The overlap figures are correlational snapshots from one vendor's dataset, not a controlled study, and they describe the dates given.
• Crawler documentation describes stated behaviour. It is not proof that a given engine honoured it on a given day, and user agents can be spoofed.
• We have not published a controlled test showing that fixing crawler access changes citation rates. Nobody we can find has either.
• Grok and Meta AI matter, particularly in India, and are absent here because xAI publishes no crawler documentation and we will not infer one.
• Checked 23 September 2026. These systems change monthly and a dated claim describes its date.